Setting_up_generator.md

Dataset setup components

In order to generate a new dataset of garments, you would need:

• Sewing Pattern Template
• Base body model
• Simulation and base rendering configuration
• (Optional) Scan imitation configuration (usually combined with Simulation props config)
• (Optional) Maya Scene for nice rendering

Examples of each component are provided in ./data_generation subfolders. Any examples can be freely reused with proper attribution given. Note, that the provided human body model is part of SMPL Body Model that is licensed under CC BY 4.0.

Preview your setup in GarmentViewer GUI

We provide simple UI for Maya for viewing a parametrized garment template, test parameter values, simulation and scan imiation processes & rendering setup.

To use it, simply run Autodesk Maya and copy the contents of garmentviewer.py to the Python scripting console.

Features:

• Loading sewing pattern spec as geometry
• Loading scene setup and simulation\scan imitation properties from external files
• Controlling parameters values (for parameters from spec) from the interface, including restoration of template state and drawing random samples
• Editing panels' global rotation and translation both from GarmentViewer or with native Maya tools
• Editing material properties for sim (native Qualoth tools)
• Running physics simulation on currently loaded garment.
• Testing the presence of collisions in garment mesh
• Visualing Segmentation labels
• Running scan imitation process
• Saving current garment with current parameter values, panel positions, garment mesh and chosen material properties with or without rendering images.

Closer look at the setup components

Pattern template

data_generation/Patterns contains exmaples of parametrized pattern templates that can be sampled to form a dataset of sewing patterns.

• Each template is a .json file that describes the structure of the base pattern and the way it could be changes (parametrization)
• docs/template_spec_with_comments.json gives a detailed description of the format used to describe pattern pemplates.

A development of a new pattern template is a mostly manual process. One should create and define a new .json file with description of the sewing pattern of base garment and specification of its parametrization.

To ease this process, we recommend to

• Use one of the existing templates as a starting point.
• Use GarmentViewer GUI at every development step. It visualizes the current pattern and allows to test parameters and their ranges.

Body model

data_generation/Bodies contain example .obj files with 3D models used to drape a garment on

• One data generation run uses only one base model (at this stage of development)
• In general, our system allows to use any kind of 3D model for sewing patterns to be draped over
  • However, our example patterns are designed to be draped on the average female body we borrow from SMPL statistical body model in T-pose: f_smpl_template.obj

Simulation config file

Any dataset_properties.json file of existing datasets could be used in place of Simulaiton Config file. This is an easy way to reproduce simulaiton and rendering context of existing data.

data_generation/Sim_props contains configuration for dataset simulation process.

On the high level, every config contains the following

• Name of the base 3D model (body) file to use for draping
• Render setup with the following elements:
  • a desired resulution of output images
  • (optional) color to apply to garments -- RGB in range (0, 1)
  • (optional) the name of the scene file to use, if any (More info: Maya Scene for rendering setup)
• Material properties of garment fabric & body to be used for physics simulation
• Geometry resolution multiplier (the larger -- the higher the final mesh resolution of generated garment 3D models)
• Different thresholds to control the sensitivity of simulation quality checks
  • Simulation quality checks are designed to filter out garments with failed simulations to avoid biasing the training a dataset will be used for
  • Examples of bad simulation results: skirt sliding down to the legs; heavy self-intersections, etc.

To simplify the process of choosing material configuration, GarmentViewer GUI supports export from Qualoth setting directly. You may edit those in Qualoth objects, test simulation until satisfactory results are achieved, and the chosen properties will be saved as simulation_properties.json file when saving current state from GUI.

Scan imitation configuration

Configuration files in data_generation/Sim_props also contain setups for controlling scan imitation process.

The properties define the visibility test performed on every face of the garment mesh:

• the number of rays to be shout from every face (in random directions) (test_rays_num).

  It controls the amount "noise" in the corrupted version: with smaller values, more and more visible faces will be randomly removed

• the number of those rays to be visible for a face marked as visible (visible_rays_num).

  Roughly corresponds to the resolution of capture device. Faces that are only partially observed, will be more likely to be marked visible with smaller values.

Render scene

Our system supports import of rednering scenes (collection of lights, cameras, and backdrops as Maya Binary .mb files) that can be used to create good-looking renders of the draping result of each garment. Usage of scenes is optional, and if none are provided, a simple setup with single camera will be created automatically.

How to use

• Path to the render scene files should be setup in the system.json (see Dependencies and Installation instructions)

• The name of the render scene file to use should be added to the render section of the Simulation properties config as follows:

  "render": {
      "config": {
          "scene": "scene.mb", 
  
          <-- other rendering properties --->
      }
  }
  

Elements of a render scene

• Arnold Light sources (any number)
  • The setting of the light sources will be preserved at render time.
• Cameras (any number)
  • At render time, images from all the available cameras in the scene will be created
• Stage backdrops geometry -- could be as complicated as needed
  • At least one of the elements should represent the floor area of the scene -- it should have a 'floor' in its name. The body model will be automatically placed on the 'floor' level.
• Configured shaders for all backdrops geometry, body, and garment
  • Shader for garment should contain substring 'garment' in its name.
  • Shader for body model should contain substring 'body' in its name
  • Shaders for other scene elements can be defined and named freely.

Additional conditions

• All elemetes should be grouped together with the group name containing substring 'scene' (e.g. 'render_scene')
• The body model, as specified in the config file, will be automatically aligned with the 'floor' level in its Y coordinates and with the center of the whole scene geometry in its XZ coordinates.
• The scene geometry should use centimeters as base units.
• If garment color is set in the Simulation properties config, it will be applied to the garment shader regardless of the original color value of this shader.

Example render scene

NOTE: We cannot share examples of rendering scenes as .md due to licensing conditions.

NOTE: As other dataset components, scenes can be tested with GUI garmentviewer.py in Maya.